Platform-specific expansions on HLO test (hlo_to_llvm_ir.cc)

[i-chaochen]

tensorflow/tensorflow#58441

The existing HLO tests are cuda only, we would like to add ROCm-amdgcn related tests to hlo_to_llvm_ir.cc.

But it will be great if XLA can have something like i.e., GpuCodegenTest::MakePlatformSpecificLlvm from gpu_kernel_tiling_test, that can replace HLO IR as platform-related instead of put so many platform-related hlo files.

This might need to rewrite from FileCheck or code refactory, or you guys could have a better design or alternative on it?

Thanks!